Home / Tutorials / create_training

Creating a training data set

Sample geometries, run PySCF, and write an HDF5 training database. Open this page as a rendered notebook, or download create_training.ipynb and run it locally after you install QMLearn.

Creating a Training Data Set

import os
import ase
from ase.io.trajectory import Trajectory

from qmlearn.drivers.mol import QMMol
from qmlearn.preprocessing import get_train_atoms
from qmlearn.io.hdf5 import DBHDF5
Engine properties

  1. Define the basis set, level of theory, exchange correlation functional and the total charge of your system:

    Basis set available in PySCF : https://pyscf.org/_modules/pyscf/gto/basis.html

    Level of Theory available in PySCF : https://pyscf.org/user.html

    Exchange Correlation Functional available in PySCF : https://pyscf.org/_modules/pyscf/dft/xcfun.html

basis = 'cc-pvTZ'
method = 'rks'
xc = 'lda,vwn_rpa'
charge = 0
Training set
  1. From a molecular dynamics trajectory 'h2o_vib.traj' remove similar structures using the get_train_atoms function and keep them as train_atoms
nsamples=10000
tol = 1E-3
mdtraj = 'h2o_vib.traj'

train_atoms=get_train_atoms(mdtraj, nsamples=nsamples, tol=tol)
nsamples = len(train_atoms)

WARN : Only get 36 samples at 90 step. Maybe you can reduce the 'tol'.
Reference structure
  1. Using train_atoms define the reference structure.
refqmmol = QMMol(atoms = train_atoms[0],  method = method, basis=basis, xc = xc)
Properties definition

  1. Create a list prop with the properties you wish to store after running PySCF engine using qmmol.run().

  2. Iterate over each geometry in train_atoms, run PySCF engine using qmmol.run() and keep the results in properties dictionary.

prop = ['vext', 'gamma', 'energy', 'forces']
properties= { k: [] for k in prop}

for atoms in train_atoms:
    qmmol = refqmmol.duplicate(atoms)
    qmmol.run()
    properties['vext'].append(qmmol.engine.vext)
    properties['gamma'].append(qmmol.engine.gamma)
    properties['energy'].append(qmmol.engine.etotal)
    properties['forces'].append(qmmol.engine.forces)
Database init
  1. Initialize the hdf5 database name and create the object db.
dbfile = os.path.splitext(mdtraj)[0]+'_QML_set.hdf5'
db = DBHDF5(dbfile, qmmol=refqmmol)
Writing Database
  1. Write the database object db and close it.
db.write_qmmol(refqmmol)
db.write_images(train_atoms, prefix='train')
db.write_properties(properties, prefix='train')
print(db.names)
db.close()

['rks', 'rks/qmmol', 'rks/train_atoms_36', 'rks/train_props_36']
Check Database
  1. Check the database by oppening and reading it.
db = DBHDF5(dbfile, qmmol=refqmmol)
data = db.read_properties(db.get_names('*/train_props_*')[0])
db.close()
data['energy'][0]

-76.09382403593892